What are the environmental costs of AI?

• AI training requires vast amounts of data, computing power
• Tech firms face pressure over environmental impacts of AI
• Radical approaches required from industry, experts say

As artificial intelligence advances at breakneck pace, the world's biggest tech companies face growing pressure to address the environmental and climate impacts of AI - which requires large amounts of electricity and water to run.

Tech giants such as Amazon, Google and Microsoft have pledged to tackle the climate crisis, yet green experts say the sector is not doing enough to mitigate the rising consumption of resources.

Electricity consumption from data centres, AI and cryptocurrency could double by 2026, according to a new report from the International Energy Agency.

Here's how the growth of AI and tech firms' resource use are raising concerns about energy, water shortages, and global warming:

How is AI requiring ever-more energy?

Developing AI has come at greater computing expense, increased by the development of new consumer products.

In February 2023 Google announced an AI tool called Bard, which it expects to reach billions of users. Microsoft is adding a button for its 'Copilot'-branded AI to Windows keyboards.

"Search tools like Google could see a tenfold increase of their electricity demand in the case of fully implementing AI in it", the IEA writes.

Amazon, Microsoft, Google, and Meta more than doubled their combined energy use between 2017 and 2021, rising to around 72 terawatt-hours (TWh) in 2021, according to the International Energy Agency (IEA).

That is equivalent to approximately one quarter of all the energy used by the United Kingdom in 2022.

Google's 'Scope 2' greenhouse gas emissions - those coming from the buying of power and heat - rose 37% from 2021 to 2022.

The information and communication technology (ICT) sector emits between 2% and 4% of all the carbon emissions produced each year, according to 2020 research from Lancaster University.

"Whatever way you look at it, if the sector wants to keep pace with the wider climate commitments of the Paris Agreement, it will need to reverse course and reduce emissions," Anne Pasek, a technology and climate researcher at Trent University in Canada, wrote in a recently-published online magazine about data centres.

"That requires changing norms and habits — probably both for consumers and industry players."

How much water are tech giants using as AI evolves?

The development of AI by the tech sector has led to a huge rise in water use. Training models involves feeding vast amounts of data into algorithms called Large Language Models (LLMs), which are computationally intensive and need powerful hardware.

In Google's latest Environmental Report, the company consumed 5.6 billion gallons of water in 2022, an increase of 1.3 billion since 2021, and 2.2 billion since 2020. The 2022 usage equates to about 10 days worth of water for the entire city of London.

Microsoft consumed nearly 1.7 billion gallons of water from its operations in 2022, an increase of about 34% from 2021, the company wrote in its latest sustainability report.

Research from University of California, Riverside in April found that just training GPT-3, the language model that was used to power OpenAI's ChatGPT, in Microsoft's U.S. data centres consumed 700,000 liters (154,000 gallons) of clean freshwater.

"ChatGPT needs to 'drink' a 500ml bottle of water for a simple conversation of roughly 20-50 questions and answers, depending on when and where ChatGPT is deployed", the researchers wrote.

"All these numbers are likely to increase by multiple times for the newly-launched GPT-4 that has a significantly larger model size," they said

The tech industry's intensive water use comes as global demand soars and supplies dwindle.

The United Nations has predicted that the need for water will exceed supply by 40% by 2030, and estimated that the number of people in cities facing water scarcity will rise from 930 million in 2016 to between 1.7 and 2.4 billion people in 2050.

Google plans to build a data centre in Uruguay's capital Montevideo, fuelling fears about water consumption given that the country is facing its worst drought in 74 years.

A spokesperson for the search giant told Context that the "data centre project is still in the exploratory phase, and our technical team is actively working with the support of national and local authorities to get this right".

By 2030, the company aims to replenish 120% of the freshwater it consumes by investing in projects which restore or conserve resources in watersheds, like rivers or lakes. As of 2022, it was replenishing only 6%, according to its report.

What are the possible solutions?

AI could see a reduction in its use after companies have finished experimenting with new tools such as ChatGPT, said Ayse Coskun, an engineering professor at Boston University.

They could then determine which areas require complex models, and where more simple ones would suffice, she said.

"People have started to think about that: 'Do I need to really throw a large hammer at this little nail, when maybe I can just use a screwdriver?'," Coskun told Context.

Efficiency improvements and regulation will also be crucial to reduce consumption across the world, analysts predict. The European Commission has implemented mandatory reporting obligations for energy use and emissions from data centres.

China requires all public organisations to be entirely powered by renewables by 2032, and the U.S. Department of Energy is funding the development of more efficient semiconductors over the next two decades.

Yet more radical approaches might be necessary to ensure companies develop in line with climate goals, experts say.

"We need to go from viewing energy efficiency (and) lower carbon footprint impact as an 'added value' to making them a first order constraint for any computer system, especially for large-scale data centers", Coskun added.

"We need accountability, more transparent reporting of carbon impacts, and more innovation to optimize energy as a whole system."

This story was updated on Wednesday, January 24, 2024 at 6:00 GMT to reflect the publication of the IEA report.

(Reporting by Adam Smith, Editing by Kieran Guilbert and Zoe Tabary.)